0009 — The phonetic encoders follow jellyfish whole

Status: accepted · Date: 2026-09-27

Context

0005 holds Soundex, NYSIIS, Metaphone and the match rating to jellyfish 1.2.1, and draws the scope of that parity short of jellyfish's whole behaviour: Metaphone is proven on real words, jellyfish's answers on degenerate letter sequences being its C implementation's, of no value to reproduce. 0007 records a second boundary as a deliberate divergence: jellyfish measures a match rating codex in UTF-8 bytes for the truncation, the length-gap check and the minimum-rating lookup, and the C# measured characters.

The pre-1.0 review of 2026-09-26 found the boundary drawn in the wrong place (#1194, #1195, #1198). The divergences were not confined to non-words: 54 of 5,402 real ASCII words differed in Metaphone (Fletcher, Mitchell, Highness), 9 in NYSIIS (Reeves, Fischer), and an apostrophe — Keats's, O'Brien — changed the Soundex, NYSIIS and match rating answers of every possessive in the dictionary, because the C# dropped what jellyfish reads. The implementation behind jellyfish 1.2.1 is Rust, not C, and its rules are visible in behaviour: a full uppercase mapping (ß is SS), a decomposition or a grapheme cluster as each encoder takes one, and every other character standing where it is.

Decision

Soundex, NYSIIS, Metaphone and the match rating reproduce jellyfish 1.2.1 on every input, not on real words only. The input is uppercased by the full case mapping; Soundex and Metaphone read its NFKD decomposition, NYSIIS and the match rating its extended grapheme clusters; a character no rule names is kept where it stands, and acts there as jellyfish's rules make it act. The match rating measures its codex in UTF-8 bytes, as jellyfish does, and Compare answers null where jellyfish answers None for a name the codex refuses.

This amends 0005's phonetics row: the Metaphone corpus is no longer the boundary of the claim, and degenerate sequences are inside it. It withdraws 0007's match rating row: byte length is no longer a divergence, so Codex("並丝七世") is jellyfish's 並丝七丝七世. The rules are reproduced from what jellyfish answers, as 0002 allows, and each is proven by its oracle, not by the source.

Consequences

• Two Unicode facts .NET does not expose are tables in Lodestar.Text: the scalars whose full uppercase differs from char.ToUpperInvariant, and the Other_Alphabetic marks char::is_alphabetic admits. Both were measured over every scalar against jellyfish 1.2.1 and Python's unicodedata 15.0; a scalar Unicode assigned after that follows the runtime's tables, and a grapheme boundary follows the runtime's segmentation, which on .NET Framework predates extended clusters.
• A differential test replaces the real-word boundary: 134,334 inputs — the dictionary and 30,000 random strings over apostrophes, spaces, ß, combining marks, emoji and letter numbers — agree with jellyfish on all four encoders, and 33,003 pairs on the comparison. The corpora gain the names and the non-ASCII cases that used to sit outside the claim.
• Metaphone's corpus, metaphone.json, is no longer special: it stays because its names cover the rules, not because it bounds what is promised.